An Active Learning Interatomic Potential For Defect-Engineered CoCrFeMnNi High-Entropy Alloy

TL;DR

This paper develops a high-fidelity Moment Tensor Potential (MTP) for the equiatomic Cantor alloy CoCrFeMnNi by combining Special Quasirandom Structure configurations, defect-perturbed states, and temperature-varied DFT data. An active-learning loop using an extrapolation-grade metric on MD trajectories systematically augments the training set, yielding a robust model (L8-RB10) that matches or surpasses MEAM in lattice, defect energetics, and phase behavior while delivering near-DFT accuracy at MD-scale speeds. The MTP is validated against DFT and experiment across properties including equation of state, vacancy formation energy, uniaxial tension, nanoindentation, and solid–liquid transitions, and demonstrates superior scalability on HPC systems. The work provides open-source access to the MTP and associated workflows, enabling rapid, large-scale screening and design of defect-engineered HEAs with high computational efficiency and predictive fidelity.

Abstract

High-entropy alloys (HEAs) exhibit exceptional properties arising from a combination of thermodynamic, kinetic and structural factors and have found applications in numerous fields such as aerospace, energy, chemical industries, hydrogen storage, and ocean engineering. However, a large compositional space remains to be explored. Unlike conventional approaches, computational methods have shown accelerated discovery of novel alloys in a short time. However, the lack of interatomic potentials have posed a challenge in discovering new alloy compositions and property measurements. In the present work, we have developed a Moment Tensor Potential (MTP) trained by Machine Learning based approach using the BFGS unconstrained optimization algorithm for the CoCrFeMnNi High-entropy alloy. Our training set consists of various defects induced configurations such as vacancies, dislocations and stacking-faults. An active learning scheme to re-train the potential was undertaken to dynamically to add training data upon encountering extrapolative configurations during non-equilibrium simulations. A thorough investigation of the error metrics, equation of state, uniaxial tensile deformation, nano-indentation and solid-liquid interface stability for this alloy was carried out, and it is seen that the MTP potential outperforms the popular Modified Embedded Atom Method (MEAM) potential on physical properties prediction. The accuracy and high computational speed are discussed using scaling performance. The potential is prepared for public use by embedding it into the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) code.

Figures (8)

• Figure 1: Schematic overview of the MLIP workflow.
• Figure 2: Validation of the trained MTP on 100 distinct SQS structures. (a)predicted against actual energy in units of eV per structure in TS1. (b),(c),(d) predicted against actual forces per atom in eV/Å for all the structures in TS1 along x,y and z axis respectively.
• Figure 3: Validation of the trained MTP on 100 distinct structures sampled at equal intervals from NVT tajectory simulated on LAMMPS. (a) predicted against actual energy in units of eV per structure in TS2. (b),(c),(d) predicted against actual forces per atom in eV/Å for all the structures in TS2 along x,y and z axis respectively.
• Figure 4: DFT, MEAM and MLIP Energy - Volume EOS comparisons for the HEA. Energy deviations from the minimum energy configuration ($E - E_{min}$) are plotted as a function of the lattice parameter. Dashed curves represent Birch Murnaghan fits to DFT, MEAM and our trained MLIP results.
• Figure 5: Comparison of the uniaxial tensile deformation between the trained MTP and MEAM in blue and red curves respectively. The Young's modulus indicated in the plot is extracted from the linear portion of the curve, under elastic region indicated in grey.